Research map / Vision-language and multimodal / Multimodal LLMs

IMAIA: Interactive Maps AI Assistant for Travel Planning and Geo-Spatial Intelligence

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Presents IMAIA, a modular geospatial assistant combining Maps Plus (quadkey-based visual prompting over maps), PAISA (camera-grounded place search and bearing-aware navigation), and a distilled Florence-2 spatial reasoning module, coordinated by a multi-agent orchestrator.

Problem

Map applications are point-and-click and offer little support for view-conditioned map queries or for connecting camera views to nearby places and last-100-meter navigation.

Approach

Converts the map viewport into a quadkey-indexed grid used as a structured visual prompt for an MLLM, grounds camera images to places by fusing CLIP similarity with distance and heading features ranked by XGBoost, and distills GPT-4o spatial annotations into a small Florence-2 model for fast scene relation extraction.

Key results

• Maps Plus reaches 89.83% POI detection accuracy versus 39.30% to 42.74% for single-model and location-augmented baselines.
• Distilled Florence-2 spatial model reaches 84% accuracy judged by o1, versus 27% for Florence-VL 8B, with 1.7s per query vs 12.4s for an agent pipeline (7.3x speedup).
• Human-centered guidance reduces walking time to 63.5% of TBT in turn-required scenarios and 32.1% in directly visible scenarios (10 scenarios).

Surprising findings

• Adding coordinates or verbose location to the MLLM baseline barely improves POI detection (41.46% and 42.74% vs 39.30%).

Limitations

• Spatial-module evaluation relies on an LLM-as-judge (o1) and a small 400-image test set.
• Navigation user study is small (10 scenarios) with informal feedback from about 15 users.
• Synthetic map query benchmark is generated by GPT-4o and covers only 10 US cities.

Experimental setup

Datasets: 40k street-view images 4,300 synthetic map queries over POIs in 10 US cities 500 image queries with annotated venues (ranker training) 400 street-view images (spatial evaluation test set)

Benchmarks: POI detection accuracy on 4,300 synthetic map queries venue ranking Precision/Recall@Top-1/Top-3 street-view spatial understanding test set (400 images) last-100-meter walking time on 10 scenarios

Metrics: accuracy precision@k recall@k inference time walking time relative to TBT

Baselines: Single Model Model + Location Model + Verbose Location Distance-sorting Similarity-sorting Florence-VL 8B ASMv2 turn-by-turn navigation

Builds on: CLIP Florence-2 YOLO-World Depth Anything V2 XGBoost

Closest papers by meaning

• MapEval: A Map-Based Evaluation of Geo-Spatial Reasoning in Foundation Models (ICML 2025 · similarity 0.87)
• TraveLLaMA: A Multimodal Travel Assistant with Large-Scale Dataset and Structured Reasoning (AAAI 2026 · similarity 0.86)
• Perceive, Interact, Reason: Building Tool-Augmented Visual Agents for Spatial Reasoning (NeurIPS 2026 · similarity 0.86)
• Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.85)
• SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning (NeurIPS 2026 · similarity 0.84)
• ERGeoBench: A Comprehensive Benchmark for Embodied Reasoning and Geo-localization in Multimodal Large Language Models (ICML 2026 · similarity 0.84)